Five million industrial robots are now working in factories worldwide. That is the number worth carrying into today’s AI news today: not a projection, a demo or a humanoid doing gymnastics, but an installed base that grew 9% during 2025. More than 600,000 new units went in during the year, and China accounted for 59% of them.

The direct answer to the morning’s main question is that physical automation has crossed a scale threshold, but “robot” still mostly means a programmable industrial machine—not a general-purpose android. The rest of the news shows the software layer catching up. Researchers are wrapping robot policies in executable programs, professional publishers are embedding AI inside tax work, and financial-data companies are combining to own more of the decision chain. The competitive edge is moving from clever outputs to systems that can be measured, supervised and integrated with actual work.

AI news today: five million robots are operating in factories

The International Federation of Robotics’ new World Robotics 2026 figures put the global operational stock at a record five million units in 2025, up 9%. Annual installations rose 11% to more than 600,000. China installed 354,000—20% more than a year earlier—and Chinese suppliers captured 55% of their home market. The United States moved ahead of Japan into second place with almost 38,500 installations, up 12%.

Europe was the awkward part of the chart: installations fell 8% in Germany, 11% in Italy, 8% in France and 15% in Spain. IFR forecasts 655,000 global installations in 2026, but that is a forecast, not an order book. Also keep the categories straight. A separate Reuters report on IFR data counted only about 7,000 humanoid robots sold in 2025. The revolution on the factory floor remains mostly specialized, bolted down and economically justified one task at a time.

HarnessPAI gives robot policies a program they can inspect

A new research preprint introduces HarnessPAI, a framework that places executable code around an existing robot action model. During a rollout, a fixed program guides and checks execution; between rollouts, feedback revises that program and turns failures into reusable skills. The authors report a 61.6-point improvement over the π0.5 baseline on LIBERO-PRO and a 27.2-point gain over WorldDreamer on RoboCasa tasks, without retraining the underlying model.

Those are paper results, not field reliability. Still, the design is useful: separate the learned action policy from an inspectable control layer. Builders should watch for independent reproduction, hardware safety tests and performance on long, messy tasks. Five million factory robots create an enormous market for better orchestration—but also an enormous blast radius for code that mistakes a benchmark for a shop floor.

Huawei takes its work agent across the HarmonyOS device stack

Huawei launched Xiaoyi Work for HarmonyOS phones, tablets and computers, according to AIBase’s September 24 report. The assistant is positioned for office work, coding and creative tasks, with cross-device continuity and a 31-day introductory allocation of 1,000 AI points. An earlier TechNode report documented the beta opening on September 16.

The interesting part is not another assistant icon. It is control of the operating-system surface across three device classes. That can reduce friction between a request, the files it needs and the app where work happens. It also concentrates permissions. Huawei’s broad capability claims still need independent testing, especially for coding quality, data handling and what the “points” buy after the trial.

Wolters Kluwer puts Libra AI inside Belgian tax work

Wolters Kluwer Legal & Regulatory announced that it is integrating Libra AI workflows into monKEY, its platform for Belgian tax professionals. The company announcement distributed by Business Wire is more consequential than its narrow geography suggests. Domain AI is moving into the place where professionals already research, calculate and document decisions, rather than asking them to shuttle sensitive facts into a general chatbot.

That distribution advantage can matter more than a marginally better model. The unanswered questions are the important ones: which sources ground an answer, how changes in Belgian tax law propagate, what reviewers can audit, and where liability sits when generated work is wrong. Integration removes clicks; it does not remove professional responsibility.

CRISS tests citation-grounded AI for cancer registrars

Another September 24 preprint describes CRISS, a retrieval-augmented assistant for oncology data specialists who interpret frequently changing cancer-coding and staging standards. Across the authors’ evaluation, RAG configurations produced substantially higher grounding scores than models answering without retrieval: 0.62, 0.56 and 0.59 across easy, medium and hard questions, versus 0.29, 0.26 and 0.29.

The study uses an LLM-as-judge protocol and has not yet cleared peer review, so those scores are evidence from one research setup, not clinical validation. The useful design choice is explicit: citations support assistance, while a human retains the final abstraction decision. In high-stakes knowledge work, “find the governing passage and show it” is a better product requirement than “sound certain.”

Glass Lewis and Clarity AI combine data, governance and voting workflows

Glass Lewis and Clarity AI announced a completed combination that brings sustainability data and AI analytics together with governance research, engagement and proxy-voting infrastructure. The companies say the merged business has more than 900 employees across 20 offices; Glass Lewis serves over 1,300 investment managers and pension funds. Financial terms were not disclosed.

This is a data-and-workflow consolidation story with AI attached, not proof that algorithms will make better votes. Joining portfolio analysis to stewardship can reduce handoffs and create a stronger evidence trail. It can also make one provider more central to consequential decisions. Watch how the company protects research independence, explains model-assisted recommendations and manages conflicts between speed, client customization and consistent methodology.

Watch & Learn

What is a Digital Twin? from MATLAB explains how a live virtual representation differs from a static simulation and why sensor data matters. It is a compact introduction for operations leaders and builders connecting AI to physical equipment. Set aside about five minutes, then ask which single asset in your organization has enough reliable data to justify a twin.

AI, Translated: embodied AI

Embodied AI is intelligence that perceives and acts through a physical system such as a robot, vehicle or smart machine. A factory arm using cameras to locate differently positioned parts and adjust its grip is a concrete example; its decisions change the real world, not just a document. You should care because physical action makes latency, safety, sensing and recovery as important as model accuracy. A confident wrong answer is annoying; a confident wrong movement can stop a line—or injure someone.

Try This Today: turn maintenance logs into a failure-mode brief with Claude Cowork

Goal: find one preventable source of downtime in 10 minutes.

1. Place a sanitized CSV of equipment incidents and a short field-definition note in a new folder.
2. Open Claude Cowork, give it access only to that folder, and ask for a frequency-versus-severity table plus missing-data flags.
3. Require a second pass that challenges the top pattern and lists what a technician must verify before action.

Copy-ready prompt: “Analyze these maintenance logs. Group recurring failure modes, show count, downtime and evidence rows, and separate recorded facts from inference. Challenge the apparent top cause with two alternative explanations. Finish with one low-risk inspection a technician can perform; do not recommend changing machinery settings.”

Anthropic says Claude Cowork is available in Claude Desktop on paid Pro, Max, Team and Enterprise plans; Linux support is beta. Remove personal data and secrets, narrow folder access, and have a qualified human validate the recommendation.

One thing to remember

AI becomes economically important when it survives contact with workflow, evidence and hardware. Five million factory robots make the point rather efficiently: operational scale belongs to systems that can repeat useful work, not merely describe it.

For continuity, read yesterday’s brief on Claude’s CRISPR-like discovery and AI crossing into higher-stakes workflows, or browse the AI & Tech Morning Brief archive.